▸ guide
Best Cheap LLMs in 2026
Under $1/M Tokens
Paid models above $0 and no more than $1 per million input tokens, ranked by benchmark score per dollar. Real data — no sponsored rankings.
▸ top budget models · score / dollar ratio
| # | Model | Score | Input $/M | Score/$ |
|---|---|---|---|---|
| 1 | Mistral: Mistral Nemo | 55.8 | $0.02/M | 2790 |
| 2 | Google: Gemini 2.0 Flash Lite | 81.2 | $0.07/M | 1083 |
| 3 | DeepSeek: DeepSeek V4 Flash | 81.4 | $0.09/M | 904 |
| 4 | Google: Gemma 3 27B | 65.5 | $0.08/M | 819 |
| 5 | Qwen: Qwen3 32B | 58.8 | $0.08/M | 735 |
| 6 | Google: Gemini 2.0 Flash | 65.9 | $0.10/M | 659 |
| 7 | Meta: Llama 3.3 70B Instruct | 59.8 | $0.10/M | 598 |
| 8 | Qwen: Qwen3 30B A3B | 53.8 | $0.12/M | 448 |
| 9 | Mistral: Mistral Small 4 | 60.6 | $0.15/M | 404 |
| 10 | Meta: Llama 4 Maverick | 76.6 | $0.20/M | 383 |
| 11 | DeepSeek: DeepSeek V3.2 | 80.5 | $0.21/M | 375 |
| 12 | Google: Gemini 2.5 Flash | 74.6 | $0.30/M | 249 |
score/$ = overall score ÷ input price per million tokens · score date unavailable · methodology
▸ price tiers at a glance
Free / $0/MExcluded from this ratio table because division by zero makes score-per-dollar meaningless. See the free page instead.
≤$0.10/MVery low input cost. Compare task accuracy before using it for high-volume work.
$0.10–$0.50/MBudget-priced API entries. The table shows whether higher score justifies the additional spend.
$0.50–$1/MUpper end of this budget set. It wins only when the measured score compensates for its cost.
▸ tasks where cheap models shine
Classification & tagging
Structured extraction (JSON output)
Summarization and rewriting
Simple Q&A over context
Code review and small edits
Embedding generation
▸ tasks that need frontier models
Complex multi-file code generation
Long-horizon planning (agents)
Mathematical proof & olympiad-level reasoning
Novel research synthesis
High-stakes decisions requiring high recall
▸ frequently asked
What is the best cheap LLM under $1 per million tokens?
Mistral: Mistral Nemo leads measured paid models priced above $0 and no more than $1 per million input tokens, at 2790 score points per input dollar.
How do I measure LLM value for money?
Divide the benchmark score by the cost per million tokens. A model scoring 65 at $0.50/M is better value than one scoring 75 at $3/M for cost-sensitive applications. However, also factor in task-specific accuracy — a cheap model that fails at your task has infinite effective cost.
Are cheap LLMs reliable for production use?
Benchmark quality and price do not prove production reliability. Check provider uptime, limits, data handling, and task-specific evaluation before deployment.